Zero Framework Cognition: A Way to Build Resilient AI Applications (steve-yegge.medium.com)

Zero Framework Cognition (ZFC) is an architectural pattern and onboarding discipline for AI-enabled applications that insists orchestration remain a “thin, safe, deterministic shell” while delegating all decision-making and semantic reasoning to external models. The author formalized ZFC after repeatedly seeing agent code bake brittle heuristics—keyword matching, regex hacks, local ranking or fallback rules—into client logic. ZFC’s concrete rules let the framework handle only IO, plumbing, schema and safety checks, policy enforcement, mechanical transforms, state tracking and typed errors; anything that requires classification, selection, planning, semantic analysis, or quality judgment must be outsourced to a model. The prescribed flow is: gather raw context, call the model for decisions, validate structure, and execute mechanically. The significance is practical and architectural: following ZFC reduces brittleness, edge-case failures, and maintenance pain by keeping cognitive complexity inside AI workers (the “smart endpoints”) rather than in fragile client heuristics, echoing Fowler’s “Smart Endpoints, Dumb Pipes” and Karpathy’s Software 2.0 ideas. The main trade-off is cost—more model calls—but ZFC makes cost optimization straightforward via a model pyramid (route low-cognition tasks to cheap models, high-cognition tasks to expensive ones) and explicit AI-driven decomposition. For orchestration-heavy, continual-workflow apps, ZFC promises greater resilience and easier debugging; it’s less clear whether it’s universally optimal, but it’s a practical blueprint for turning brittle hacks into model-driven decisions.